Ayoub
Cherguelaine

Problem: Document-AI training needs varied business layouts with aligned text, field values, and coordinates for evaluating extraction.

Solution: Built a seeded pipeline that renders documents for 60 synthetic companies, checks text and financial totals, and exports gold fields, word boxes, and degraded scans.

Problem: Regulations, rules, ministry systems, and salary tables arrive in different formats. Collecting them is only the first step: extraction, provenance, storage, and failed processing stages also need to be traceable.

Solution: Built four source pipelines with scheduled execution, resumable stages, Arabic OCR, and PostgreSQL/Neo4j loading, alongside a Next.js monitor for inspecting documents, tracking costs, and managing operations.

Problem: Batch-style legal analysis flows produced delayed final answers with little transparency into intermediate extraction, retrieval, and reasoning quality.

Solution: Implemented a FastAPI `/process_doc_stream` pipeline backed by `StreamingLegalCaseProcessor`, using SSE events, token-aware chunking (`cl100k_base`), Pinecone similarity lookup, multithreaded case comparison, and incremental prediction refinement.

Problem: Passengers need a clear way to assess disrupted flights, bring together tickets and airline messages, and understand which facts and legal sources support a claim.

Solution: Built a Next.js case room and FastAPI backend with a guided claim form, local PDF extraction and Tesseract OCR, bundled airport lookup, and two-stage LLM assessment. The deeper review retrieves passages from 36 bundled legal records and returns reasons with validated citation IDs and exact archived quotes.

Problem: Exploring unfamiliar tables takes SQL knowledge, while a fluent AI answer alone does not show which data or queries support its claims.

Solution: Built a source-aware analyst that inspects schemas, samples tables, executes guarded read-only SQL, and answers from observed results. Users can import datasets, select tables, follow tool activity, and request written analysis.

Problem: News categorization needed a reusable model that could score candidate topic labels in a standard inference workflow.

Solution: Fine-tuned facebook/bart-large-mnli on CNN articles and published the model, tokenizer, and a Transformers inference example.